Join our Newsletter — 33% off our NHI Course
Home› Guides› EU AI Act Article 6(3) Guide: When an…
Guide AI Regulation & Law

EU AI Act Article 6(3) Guide: When an Annex III System Is Not High-Risk and How to Document It

PDF for newsletter subscribers · free to read online

← All guides
By NHI Mgmt Group, under the editorial direction of Lalit Choda Updated 18 November 2026 13 min read
On this page

Article 6(3) of the EU AI Act lets a provider conclude that an AI system listed in Annex III is not high-risk, but only on four narrow grounds, never where the system profiles people, and only with a documented assessment that a market surveillance authority can challenge. This guide explains the test in the text of Regulation (EU) 2024/1689, how to apply each condition to real systems including AI agents, what the assessment record should contain, the registration that still applies, and what happens if an authority disagrees.

For the full list of Annex III use cases and the obligations that follow when a system is high-risk, see the EU AI Act High-Risk AI Systems Guide. This page covers only the derogation and how to evidence it.

This guide is not legal advice. Status as of 17 Nov 2026: it reflects Regulation (EU) 2024/1689 as amended by Regulation (EU) 2026/1744 (the Digital Omnibus on AI). Classification decisions should be reviewed by qualified counsel.

Key takeaways

  • Article 6(2) makes Annex III systems high-risk by default. Article 6(3) is a derogation: a system is not high-risk where it does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons, including by not materially influencing the outcome of decision making.
  • The derogation applies where at least one of four conditions is met: a narrow procedural task, improving the result of a completed human activity, detecting decision-making patterns or deviations without replacing human assessment, or a preparatory task.
  • An Annex III system that performs profiling of natural persons is always high-risk. No condition overrides that.
  • The provider must document the assessment before placing the system on the market or putting it into service, register under Article 49(2), and hand the documentation to national competent authorities on request (Article 6(4)).
  • Authorities can re-evaluate the classification under Article 80, and misclassification to circumvent the high-risk requirements can be fined under Article 99.
  • Chapter III Section 1, which contains Article 6, applies to Annex III systems from 2 December 2027 after the Digital Omnibus on AI amended Article 113.

Article 6(3) at a glance

ElementWhat the text saysPractical meaning
Starting pointAnnex III systems are high-risk under Article 6(2)You need a reason to step out, not a reason to step in
Core testNo significant risk of harm to health, safety or fundamental rights, including by not materially influencing decision outcomesThe question is influence on decisions about people
Condition (a)Intended to perform a narrow procedural taskFormatting, routing, converting, deduplicating
Condition (b)Intended to improve the result of a previously completed human activityPolishing a human's finished work
Condition (c)Detects decision-making patterns or deviations, not meant to replace or influence the completed human assessment without proper human reviewAfter-the-fact consistency checks
Condition (d)Preparatory task to an assessment relevant to an Annex III use caseGathering or organising material before a human assesses
OverrideAlways high-risk where it performs profiling of natural personsScoring or evaluating individuals ends the analysis
ProcessDocument before market or service, register, provide on requestArticle 6(4) and Article 49(2)
ChallengeAuthority re-evaluation and finesArticle 80 and Article 99

Scope: who uses Article 6(3) and when

The derogation is the provider's call. Article 6(4) is addressed to "a provider who considers that an AI system referred to in Annex III is not high-risk". A deployer cannot declare a provider's system non-high-risk. A deployer can, however, become a provider under Article 25, for example by changing the intended purpose of a system so that it falls within Annex III, and then the classification question becomes its own. The EU AI Act Provider vs Deployer Obligations Guide explains those triggers.

The derogation applies only to Annex III systems under Article 6(2). It does not apply to Annex I product systems under Article 6(1), where the classification follows the product's third-party conformity assessment.

On timing, Regulation (EU) 2026/1744 amended Article 113 so that Chapter III Sections 1 to 3 apply from 2 December 2027 for Article 6(2) and Annex III systems, and from 2 August 2028 for Article 6(1) and Annex I systems. Article 6(5), which provides for Commission guidelines on the practical implementation of Article 6, is excluded from that deferral. Check the Commission's AI Act pages for the current status of its classification guidelines and their list of practical examples before finalising an assessment; they were not reviewed for this page.

The core test: significant risk and material influence

The first subparagraph of Article 6(3) states the principle: an Annex III system is not high-risk where it does not pose a significant risk of harm to the health, safety or fundamental rights of natural persons, including by not materially influencing the outcome of decision making. The four conditions in the second subparagraph are the routes by which that is shown.

Two readings help keep the test honest:

  • Influence is judged by effect, not by labels. A system described as "decision support" that ranks candidates, and whose top ranking is accepted in most cases, materially influences the outcome. Measure acceptance rates and how often humans depart from the output.
  • Intended purpose governs. Each condition begins "the AI system is intended to". The assessment must be anchored in the intended purpose stated in the instructions for use and marketing, and the system must be designed so that it cannot easily be used for more. If your sales material promises automated shortlisting, a "preparatory task" assessment will not survive.

Condition (a): narrow procedural task

A system intended to perform a narrow procedural task meets condition (a). Typical examples are systems that convert unstructured documents into structured fields, classify incoming documents into categories for routing, detect duplicates among applications, or check that a submission is complete. The system acts on the form of the material, not on the merits of the person.

Questions to answer in the record:

  • What exactly is the task, in one sentence, and is it procedural rather than evaluative?
  • Can an error affect the outcome for a person? For example, a routing error that sends an application to a rejection queue is not purely procedural.
  • Is the task narrow, with a bounded input and output, or open-ended?

Condition (b): improving a completed human activity

A system intended to improve the result of a previously completed human activity meets condition (b). The human has already done the work and made the judgement; the system refines it. Examples include improving the language of a document a human has drafted or checking a human-written decision letter for clarity.

The record should show the human activity is complete before the system acts, and that the system's change does not alter the substance of the human's judgement. A system that "improves" a human assessment by adjusting a score has moved beyond condition (b).

Condition (c): detecting patterns or deviations

Condition (c) covers a system intended to detect decision-making patterns or deviations from prior decision-making patterns, and not meant to replace or influence the previously completed human assessment without proper human review. An example is a tool that reviews a set of completed grading or claims decisions and flags ones that look inconsistent with the decision maker's usual pattern.

The safeguard is in the text: the system must not replace or influence the completed assessment without proper human review. Document who reviews flagged items, with what authority, and evidence that flags do not automatically reverse or reopen decisions.

Condition (d): preparatory tasks

Condition (d) covers a system intended to perform a preparatory task to an assessment relevant for the purposes of the Annex III use cases. Examples include indexing and searching files, translating documents, or summarising a case file so that a human assessor can work from it.

This is the most stretched condition in practice. Preparation can shade into assessment: a summary that omits material facts, or a retrieval system that decides which evidence the assessor sees, can materially influence the outcome. The record should explain why the preparation is neutral, how completeness is checked, and whether assessors still have access to the full underlying material.

The profiling override

The final subparagraph of Article 6(3) states that, notwithstanding the first subparagraph, an Annex III system shall always be considered high-risk where it performs profiling of natural persons. Profiling takes its meaning from the GDPR: automated processing of personal data to evaluate personal aspects of a person, in particular to analyse or predict aspects such as work performance, economic situation, health, preferences, interests, reliability, behaviour, location or movements.

Apply the override first, before the four conditions. If the system scores, ranks, segments or predicts anything about individual people within an Annex III area, stop: it is high-risk. Many systems that look like condition (a) or (d) tools do profile, for example a triage system that routes job applications by predicted suitability.

Applying the test to AI agents and multi-step systems

Agentic systems complicate classification because one agent can perform several tasks with different characters. A recruitment assistant that extracts CV fields (procedural), drafts interview questions (preparatory) and then recommends a shortlist (evaluative, likely profiling) is one system with one intended purpose. Assess the system as placed on the market, not the most benign sub-task.

Controls that support a defensible classification:

  • Constrain the tools. If the assessment depends on the agent not scoring people, ensure it has no tool, data scope or credential that would let it do so. Tool grants and data access are part of the evidence.
  • Constrain the outputs. Use output schemas that cannot carry a ranking or score where the assessment says none is produced.
  • Watch for scope creep. Treat a new tool, connector or prompt that changes what the system does as a trigger to repeat the Article 6(3) assessment.
  • Log what it actually did. Logs showing the system's actual behaviour are the best answer to an authority's Article 80 evaluation.

Documenting the assessment and registering

Article 6(4) requires a provider who considers that an Annex III system is not high-risk to document its assessment before the system is placed on the market or put into service. The provider is subject to the registration obligation in Article 49(2), and on request of national competent authorities must provide the documentation of the assessment.

A good assessment record contains:

  1. System identification: name, version, provider, intended purpose as stated in the instructions for use.
  2. The Annex III point the system would otherwise fall under, and why.
  3. The profiling analysis, with a reasoned conclusion that the system does not perform profiling of natural persons.
  4. The condition or conditions relied on, with a factual description of how the system meets each.
  5. The material influence analysis: how outputs are used, human review steps, measured acceptance and override rates where available.
  6. Design controls that keep the system within the assessed purpose: tool scopes, output schemas, access controls.
  7. Reassessment triggers and the date of next review.
  8. Approver, date and version history.

Registration under Article 49(2) is a separate step in the EU database. The Digital Omnibus on AI did not remove the registration duty; it streamlined the Annex VIII content, deleting points 7 and 9 of Annex VIII Section B. Check the current Annex VIII Section B list when preparing the entry.

Enforcement: Article 80 re-evaluation and fines

Article 80 gives market surveillance authorities a specific procedure for systems that providers have classified as non-high-risk under Article 6(3). Where an authority has sufficient reason to consider that such a system is in fact high-risk, it evaluates the system. If it finds the system is high-risk, it requires the provider to take the necessary actions to bring the system into compliance and corrective action for systems already on the market. Article 80(7) provides that providers are subject to fines under Article 99 where the authority establishes that the system was misclassified as non-high-risk to circumvent the high-risk requirements.

Under Article 99(4), non-compliance with provider obligations under Article 16 attracts fines of up to EUR 15,000,000 or 3% of worldwide annual turnover, whichever is higher, with the lower figure applying to SMEs (Article 99(6)) and, after the Digital Omnibus, to small mid-cap enterprises (Article 99(6a)). Supplying incorrect, incomplete or misleading information to authorities, for example an assessment record that misdescribes the system, attracts up to EUR 7,500,000 or 1% (Article 99(5)). The EU AI Act Penalties and Enforcement Guide sets out the wider regime.

What it means for security and identity teams

Classification looks like a legal exercise, but much of the evidence is technical. Security and identity teams can provide:

  • An inventory of AI systems with their data sources and the identities they use, so legal can see which systems touch Annex III areas.
  • Proof that a system cannot reach data or tools that would let it profile people, through scoped service accounts and agent tool permissions.
  • Change detection on model, prompt and tool configuration, feeding reassessment triggers.
  • Logs showing how outputs are used and how often humans override them.

Even where a system is validly non-high-risk, other duties can still apply, including Article 50 transparency and Article 4 AI literacy measures. See the EU AI Act Article 50 Transparency Obligations Guide.

Common failure patterns

  • Applying the four conditions before checking for profiling, and missing that the system scores individuals.
  • Assessing the most benign sub-task of an agent rather than the system's intended purpose as a whole.
  • Marketing material or instructions for use that describe the system as making recommendations, contradicting a preparatory-task assessment.
  • Writing the assessment after launch, when Article 6(4) requires it before placing on the market or putting into service.
  • Forgetting the Article 49(2) registration because the system is "not high-risk".
  • No reassessment when new tools, connectors or data scopes are added.
  • Deployers relying on a vendor's statement that a system is "not high-risk" without asking for the assessment or checking for Article 25 triggers in their own use.

Practitioner checklist

  • List every AI system that touches an Annex III area and record the relevant Annex III point.
  • Run the profiling check first and record the result with reasons.
  • Identify the Article 6(3) condition relied on and describe the facts that meet it.
  • Analyse material influence with data on how outputs are used and overridden.
  • Align the instructions for use and marketing with the assessed purpose.
  • Restrict tools, data scopes and output formats to what the assessment assumes.
  • Complete and approve the assessment before market placement or putting into service.
  • Register in the EU database under Article 49(2) using the current Annex VIII Section B fields.
  • Set reassessment triggers on configuration change and a scheduled review date.
  • Prepare to provide the documentation to authorities on request and to respond to an Article 80 evaluation.

Standards and references

Related NHI Mgmt Group resources: EU AI Act High-Risk AI Systems Guide · EU AI Act Conformity Assessment and Harmonised Standards Guide · EU AI Act for AI Agents · EU AI Act Compliance Guide

PDF for subscribers

Take this guide with you

Every section, table and checklist in one PDF, with links to the sources and related guides. Subscribe to our weekly NHI & AI Identity Journal and we’ll email you a download link.

Free weekly newsletter

Subscribe to the NHI & AI Identity Journal

The latest on NHI and Agentic AI security – articles, research, breaches, news and events every week.

Bonus 33% off our NHI Course when you subscribe.

PDF Downloads for Newsletter Subscribers

Get the guide

Enter your details and we’ll email you a download link.

The full guide is free to read online. The PDF is a thank you for subscribing.

Fill in all fields marked * and tick both boxes to continue.